Anthropic announced Enterprise Frontier Safeguards on September 1 — misuse detection that runs on activity data stored in the customer’s own cloud account, under the customer’s own encryption keys, with no Anthropic human review required. It directly resolves the Mythos-class 30-day data retention mandate that forced GitHub Copilot to ship Claude Fable 5 off-by-default for enterprise back in June.
Meta shipped Muse Spark 1.3 on September 2 with two configurations: the one developers can actually buy scores a tie, not a lead, and got more expensive per task despite flat per-token pricing. The higher score Meta led with is locked to a ’limited partner preview’ nobody outside Meta can access.
The independent Terminal-Bench 4.0 leaderboard now shows GPT-6 Astra at 58.2% (±2.8) and Claude Fable 5.1 at 57.9% (±3.8) — a gap smaller than either model’s own error margin. Both self-reported numbers held up under independent verification, but the ‘winner’ framing both labs used doesn’t survive contact with the error bars.
OpenAI launched GPT-6 Astra on September 3, its first model to cross the ‘Critical’ cybersecurity-capability threshold under its own Preparedness Framework — able to find unknown vulnerabilities and build exploits for well-protected systems without step-by-step guidance. It’s rolling out to every ChatGPT tier and the API within days, gated by little more than an admin checkbox and a refusal prompt.
Four Claude Code releases shipped September 1-4 (v2.1.258 through v2.1.261): a new fullscreen diff panel, org-wide managed MCP servers, a headless no-prompts mode for unattended hosts — and a fix for a bug that meant Fable 5.1’s freshly announced 75% cache-read discount wasn’t actually being applied after tool calls.
Anthropic launched Claude Fable 5.1 and Mythos 5.1 on September 1, the same day Terminal-Bench 4.0’s leaderboard finally populated with real, independently-run scores. Claude Opus 5 leads at 51.8% and Fable 5 sits second at 44.5% — a 7-point gap over third-place GLM-5.3 — while Fable 5.1’s own self-reported 55.8% isn’t on that leaderboard yet.
A newsletter claim that Anthropic and OpenAI jointly published an ‘MCP v2.0’ draft with bidirectional tool calling has spread across multiple blogs since mid-August. Direct checks of the primary spec repository and the protocol’s own roadmap post find no such draft, no OpenAI co-authorship, and a versioning scheme that doesn’t use version numbers at all.
Anthropic announced it is permanently raising Claude Code’s standard weekly usage limits by 25% starting September 14 — but that number replaces a temporary 50% boost that’s been active since May, so heavy users are actually looking at a 17% cut versus what they have today.
OpenAI announced August 28 that it will end Cursor’s direct access to its models on November 12, 2026, citing SpaceX and xAI’s history of violating its terms of service. Cursor CEO Michael Truell says OpenAI models are only ~5% of Cursor’s traffic — but the episode is a clean case study in what happens when your product is a reseller of other labs’ models.
Claude Code v2.1.248 (Aug 27) ships a –restricted flag that strips exec and WebFetch tools, refuses bypassPermissions outright, and ignores project-level settings files — plus native Windows support for cross-session SendMessage and a new /usage-credits command for Enterprise. Verified directly against Anthropic’s own changelog.